Personalized AI memory adapts with geometry for better long-term use
Memory Has Geometry: Non-Uniform Geometric Memory for Long-Horizon Personalized AI
Artificial Intelligence
Summary
Personalized AI systems often remember information as fixed points you can look up, but this doesn’t match how our experiences change over time. The authors propose viewing AI memory as a moving path through a space that changes shape depending on the user’s context and time, not just fixed snapshots. This lets the AI better understand when information should stay stable or change quickly, and how to access relevant memories based on how the user’s situation evolves.
What this means in practice
- •For personalized ai developers: Design AI systems that adapt memory retrieval dynamically based on user interactions and elapsed time for improved long-term personalization.$Commercial implications: Enables development of AI products that maintain more accurate and evolving user profiles over extended interactions enhancing user experience.
- •For interactive system designers: Create interfaces that respond to users by tracking changes in user state rather than relying on static event records.
A position paper. It proposes an approach and reports no results.
Authors
Jiahong Liu, Wenhao Yu, Zexuan Qiu, Menglin Yang, Irwin King
Abstract
Long-term memory is becoming a core substrate for personalized AI, yet most systems still represent personalization as discrete records in a largely static latent space, accessed under one global similarity notion. For data mining, this creates a mismatch: the evidence is a temporal event stream, while the dominant abstraction is a searchable record set. We argue that long-horizon personalization should instead model memory as a user-specific dynamical state space with locally heterogeneous geometry. Geometry here is a computational language, not a literal claim about cognition: it captures stable versus volatile regions, variable-rate drift, heterogeneous neighborhoods, and uncertainty about current user state. Profiles and isolated events remain useful as points, but interaction, feedback, and elapsed time induce trajectories. Memory access then becomes trajectory-conditioned reconstruction of the relevant user state, not only nearest-neighbor lookup.